agentleFS
PatrickGallucciPatrickGallucci/fabric-skills/skills/fabric-notebook-perf-remediate/SKILL.md

fabric-notebook-perf-remediate

Skill17 starsChanged 9 months agoMITView on GitHub

Diagnose and resolve performance issues in Microsoft Fabric notebooks running Apache Spark. Use when notebooks are slow, Spark jobs are timing out, sessions hit HTTP 430 throttling errors, data skew causes straggling tasks, shuffle operations are inefficient, Delta Lake tables need optimization, V-Order configuration is needed, capacity utilization is high, OOM errors occur, or Spark Advisor warnings appear. Covers Spark session tuning, Native Execution Engine, autotune, custom Spark pools, partition optimization, and Monitoring Hub diagnostics.

  • Installs packages

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run the notebook health check script to audit your current session configuration.

    line 83
  2. Distinctive:

    Stop sessions you are not actively using — default idle timeout is 20 minutes.

    line 100
  3. Distinctive:

    Check the Spark Advisor light-bulb icon for memory warnings.

    line 115
  4. Distinctive:

    Use predefined Spark resource profiles to auto-configure for your workload type:

    line 183

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Microsoft Fabric Notebook Performance remediate

Systematic toolkit for diagnosing, analyzing, and resolving performance bottlenecks in Microsoft Fabric notebooks powered by Apache Spark.

When to Use This Skill

  • Fabric notebook cells are running slowly or timing out
  • Spark jobs are being throttled with HTTP 430 errors
  • Capacity Metrics app shows high CU consumption
  • Data skew is causing unbalanced task execution
  • Shuffle operations are consuming excessive resources
  • Delta Lake tables have degraded read/write performance
  • OOM (Out of Memory) errors during notebook execution
  • Spark Advisor shows warnings or errors in cell output
  • Session startup is slow or sessions expire unexpectedly
  • Pipeline-triggered notebooks are queued for extended periods

Prerequisites

  • Workspace Admin or Contributor role in the target Fabric workspace
  • Access to the Fabric Monitoring Hub for your capacity
  • Fabric Capacity Metrics app installed (for capacity-level analysis)
  • Familiarity with PySpark or Spark SQL syntax

remediate Decision Tree

Identify your symptom and follow the corresponding workflow.

SymptomRoot Cause CategoryAction
Notebook cell runs for minutes on small dataSpark session config or query planSee Spark Session Tuning
HTTP 430 error on job submissionCapacity exhausted, concurrency limitSee Capacity and Throttling
One task takes 10x longer than othersData skewSee Data Skew Diagnosis
Write operations are slowV-Order overhead or small file problemSee Delta Table Optimization
OOM / executor lost errorsMemory pressure, partition sizingSee Memory and Partition Tuning
Session expired / timed outIdle timeout settingsSee Common Errors
Notebook queued, never startsQueue limits for SKUSee Capacity and Throttling

Quick Wins Checklist

Apply these optimizations first — they resolve the majority of performance issues.

  1. Enable Native Execution Engine (NEE) — delivers 2x–5x improvement:
spark.conf.set("spark.native.enabled", "true")
  1. Enable Autotune for adaptive configuration:
spark.conf.set("spark.ms.autotune.enabled", "true")
  1. Right-size shuffle partitions — default 200 is often wrong:
# For datasets under 1 GB
spark.conf.set("spark.sql.shuffle.partitions", "20")
# For datasets 1-10 GB
spark.conf.set("spark.sql.shuffle.partitions", "100")
# For datasets over 10 GB, leave default or increase
  1. Enable Adaptive Query Execution (AQE):
spark.conf.set("spark.sql.adaptive.enabled", "true")
spark.conf.set("spark.sql.adaptive.coalescePartitions.enabled", "true")
spark.conf.set("spark.sql.adaptive.skewJoin.enabled", "true")
  1. Use DataFrame APIs instead of RDDs — enables Catalyst optimizer and Tungsten engine.
  2. Break complex query chains into staged intermediate writes to reduce Catalyst plan complexity. See Common Errors.

Run the notebook health check script to audit your current session configuration.

Capacity and Throttling

Each Fabric SKU maps to a fixed number of Spark VCores (1 CU = 2 Spark VCores). When all VCores are consumed, new jobs receive HTTP 430 errors.

SKUSpark VCoresQueue Limit
F244
F8168
F64 / P112864
F128 / P2256128
F256 / P3512256

Resolution steps:

  1. Open the Monitoring Hub and cancel idle or unnecessary Spark sessions.
  2. Stop sessions you are not actively using — default idle timeout is 20 minutes.
  3. Reduce executor count in custom Spark pools to free VCores for parallel jobs.
  4. Enable Autoscale Billing for Spark for bursty workloads — jobs use dedicated serverless resources instead of consuming capacity CUs.
  5. For pipeline-triggered notebooks, leverage job queueing (FIFO). Queue expiry is 24 hours.

Queueing is not supported for interactive notebook jobs or Fabric trial capacities.

Memory and Partition Tuning

OOM errors typically stem from oversized partitions or insufficient executor memory.

Diagnose with Spark UI:

  1. Open the cell's Spark job progress indicator.
  2. Click Resources tab to view executor usage graph.
  3. Check the Spark Advisor light-bulb icon for memory warnings.

Tune partitions:

# Check current partition count
df.rdd.getNumPartitions()

# Repartition for parallelism (increases partitions)
df = df.repartition(200)

# Coalesce to reduce partitions (avoids full shuffle)
df = df.coalesce(50)

# Adjust max partition bytes for reads
spark.conf.set("spark.sql.files.maxPartitionBytes", "128m")

Tune task memory:

# For memory-intensive tasks causing OOM
spark.conf.set("spark.task.cpus", "2")  # More memory per task

# For CPU-bound tasks needing more parallelism
spark.conf.set("spark.task.cpus", "0.5")  # More concurrent tasks

Monitoring and Diagnostics

In-Notebook Monitoring

  • Spark job progress bar — real-time per-cell execution status
  • Resources tab — executor allocation and resource usage line chart (Spark 3.4+)
  • Spark Advisor — Info/Warning/Error recommendations per cell (expand via light-bulb icon)

Monitoring Hub

Navigate to Monitoring Hub to view all active Spark applications across your workspace. Key actions: cancel sessions, view executor count, check job duration, identify queued jobs.

Capacity Metrics App

Filter by item type (Notebook, Lakehouse, Spark Job Definition) to see CU consumption per job. Use the Multi metric ribbon chart to identify capacity spikes over time.

Formula: CU consumption = Total Spark VCores / 2 × duration

Spark Pool Configuration

Starter Pools (Default)

Session initialization in 5–10 seconds, pre-configured, no manual setup. Good for development and small workloads.

Custom Spark Pools

Configure via Workspace Settings → Data Engineering/Science → Spark Settings:

ScenarioNode SizeGuidance
Transform-heavy with shuffles and joinsLarge (16–64 cores)Maximize per-node memory
Bursty or unpredictable jobsMedium + AutoscaleLet cluster grow/shrink dynamically
Many small parallel jobsSmall/MediumUse mssparkutils.notebook.runMultiple()
Development / explorationSmall, single nodeDriver and executor share 1 VM
ML / distributed trainingMany medium/large nodesMaximize parallelism

Enable Customize compute configurations for items in Workspace Settings → Pool tab to allow per-notebook pool overrides.

Resource Profiles

Use predefined Spark resource profiles to auto-configure for your workload type:

ProfileBest For
DefaultGeneral-purpose workloads
readHeavyforSpark / ReadHeavyInteractive queries, dashboards (enables V-Order)
Write-heavyData ingestion pipelines (V-Order disabled by default)

Library Management Impact

Library installation in Fabric environments takes 5–15 minutes during publishing. For interactive development, use inline installation (%pip install) to avoid environment republish delays. However, inline commands are turned off by default in pipeline runs due to dependency tree instability.

MethodSession ImpactPipeline Safe
Environment librariesNone (pre-installed)Yes
Inline %pip installCurrent session onlyNo
Inline install.packages() (R)Current session onlyNo

References

  • Spark Session Tuning — NEE, autotune, session configs, resource profiles
  • Data Skew Diagnosis — Identifying skew, AQE, salting, repartitioning
  • Delta Table Optimization — V-Order, OPTIMIZE, VACUUM, ZORDER
  • Common Errors — Error messages, timeouts, connectivity, excessive query complexity
  • Notebook Health Check Script — PySpark diagnostic script
  • Spark Config Template — Starter configuration cell

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.